personalized information services: the infofactory project

PersonalizedPersonalized InformationInformation ServicesServices::the infoFACTORY Projectthe infoFACTORY Project

Prof. Carlo TassoUniversity of Udine

Artificial Intelligence LaboratoryinfoinfoFACTORYFACTORY Group

PIA 2005 - UM’05, July 24 2005

28/07/2005 2

infoinfoFACTORYFACTORY

ContentsContents

• Problems and vision• The infoFACTORY idea• Intelligent Platforms for Personalized IA

on the Internet• Applications• Research activities

28/07/2005 3


GeneralGeneral ProblemsProblems

• Information Explosion• E-Contents Overload

• Strategic Value of Information & Knowledge (IST)

• Information is the ‘vehicle’ of Knowledge

• Unstructured information

28/07/2005 4

infoinfoFACTORYFACTORYProblemsProblems withwith ((unstructuredunstructured) ) InformationInformation

(and Knowledge)(and Knowledge)• Problems in searching, accessing, delivering…• Two perspectives: PULL & PUSH• Pull:

– Oversupply, overload– Miss-retrieval (common accuracy of serach engines: 20%-30%)– Untimeliness

• Push:– Miss-delivery– Oversupply– Waste– Untimeliness

• Two difficult linguistic phenomena: Polisemy, Sinonimity

28/07/2005 5


The The fundamentalfundamental limitationslimitations of of currentcurrent approachesapproachesand and toolstools

• Key-word based:NO semantic analysis of texts

• ‘One-size-fits all’ approach:all users are treated the same way,• NO comprehension of individual information

needs and preferences• Presentation and delivery are the same for all

users

28/07/2005 6


Towards a SolutionTowards a Solution

28/07/2005 8

infoinfoFACTORYFACTORYAutomaticAutomatic PersonalizationPersonalization

(in (in InformationInformation Access)Access)

Personalization means delivering:• the right content• to the right user• in the right moment and• in the most suitable way

28/07/2005 9


AdaptiveAdaptive PersonalizationPersonalization

• Adaptivity allows to automaticallytailor system behaviour and processing to the specificcharacteristics of the (observed) user behaviour

• Machine Learning, ConceptualLinguistic Analysis, KnowledgeRepresentation, User Modeling, HCI, Intelligent Agents, Semantic Web, …

28/07/2005 10


28/07/2005 11


ContentContent PersonalizationPersonalization

Personalized SELECTION

end user

selectedcontents

Personalized PRESENTATION

PersonalizedDELIVERY

contents

Poket PC

PC

WAP, GPRS

UMTS, Smart Phone

PDA

availablecontents

(from C.Tasso, P.Omero, La Personalizzazionedei Contenuti WEB, © F.Angeli, Milano, 2002.)

28/07/2005 12


ContentContent PersonalizationPersonalization

Personalized SELECTION

end user

selectedcontents

Personalized PRESENTATION

PersonalizedDELIVERY

contents

Poket PC

PC

WAP, GPRS

UMTS, Smart Phone

PDA

availablecontents


28/07/2005 13

infoinfoFACTORYFACTORYSemanticSemantic ((contentcontent--basedbased))

PersonalizedPersonalized InformationInformation FilteringFiltering

Filtering information by means of a ‘semantic’ analysis of a document and by matching itagainst an individual user profile

(Brajnik G., Guida G. and Tasso C. User Modeling in Expert Man-Machine Interfaces: A Case Study in Intelligent

Information Retrieval. IEEE Transaction on Systems, Man, and Cybernetics, 20(1), 1990, pp. 166-185)

28/07/2005 14


ourour viewview ……

“…Evaluating an interface for aninformation retrieval system can notrely only on standard precision and recall, but needs to take into account also user satisfaction…”

(Brajnik G., Mizzaro S. and Tasso C. Evaluating User Interfaces toInformation Retrieval Systems: a Case Study on User Support.

Proceedings of the 19th Annual International ACM SIGIR Conferenceon Research and Development in Information Retrieval, SIGIR '96,

Zurich, CH, August 18-22, 1996, pp. 128-136.)

28/07/2005 15

infoinfoFACTORYFACTORYThe IFT The IFT AlgorithmAlgorithm(Minio & Tasso, 1996)


Positive SampleDocuments

Negative SampleDocuments

Documents to be filtered(html, xml, pdf, postscript,

doc, text, latex, …)

PROFILEBUILDER

UserProfile

ConceptualContent of aDocument

LINGUISTICPROCESSOR

MATCHING

Evaluation of therelevance of

the document

Relevance feedbackfrom the

user

28/07/2005 16


The The UserUser ProfileProfile

(Semantic) network with lexical stemsrepresenting concepts associated toNODES and co-occurence informationattached to ARCS.

M. Minio, and C. Tasso, User Modeling for Information Filtering on Internet Services: Exploiting an Extended Version of the UMT Shell, UM96 Workshop on User Modeling for Information Filtering on the World WideWEB, Kailua-Kona, Hawaii, USA, January 1996.

Minio M. and Tasso C. IFT: una interfaccia intelligente per il filtraggio di informazioni basato su modellizzazione d'utente. Supplemento di AI*IANotizie, IX(3), 1996, pp. 21-25.

28/07/2005 17


UserUser ProfileProfile ConstructionConstruction

• Through analysis of sample texts(positive or negative examples)

• Lexical stemming (Porter’s algorithm)• Stop word list• Frequency of Concepts (lexical stems)• Co-occurence information

• No need for ad-hoc onthologies nor KB

28/07/2005 18


MatchingMatching DocumentsDocuments againstagainst ProfilesProfiles

• Weighted matching algorithm, takinginto account frequency of concepts and frequency of co-occurence of concepts

• Assumption: Relevance = Topicality

28/07/2005 19

infoinfoFACTORYFACTORYExperimentalExperimental EvaluationEvaluation

of the IFT of the IFT AlgorithmAlgorithm: : PrecisionPrecision

(Asnicar, Di Fant & Tasso, 1997)

28/07/2005 20

infoinfoFACTORYFACTORYExperimentalExperimental EvaluationEvaluationof the IFT of the IFT AlgorithmAlgorithm: Ranking : Ranking QualityQuality

(NDPM)(NDPM)

(Asnicar, Di Fant & Tasso, 1997)

Search Engines

IFT

28/07/2005 21

infoinfoFACTORYFACTORYThe TIPS Project in FP5 ISTThe TIPS Project in FP5 IST

20002000--20022002

• Innovative advanced services in electronic publishing(scholarly journals)

• Filtering service on new incoming information• Application in the field of High Energy Physics

(Brajnik G., Mizzaro S., Tasso C., Venuti F.,Strategic Help in User Interfaces for Information Retrieval,

Journal of the American Society for Information Scienceand Technology 53(5), 2002, pp. 343-358)

(Mizzaro and Tasso, 2002)

tips.sissa.it

28/07/2005 22


The TORII The TORII PortalPortal

The infoinfoFACTORYFACTORY Project

28/07/2005 25


The Vision:

personalized information services can help reducing information overload and

can overcome some limitationsof current approaches

28/07/2005 26


The infoinfoFACTORYFACTORY Concept

• Delivering personalized information servicesstarting from any digital information source (E-Content)

• Adaptive personalization by means of content-based individual profiles

• Overcoming information overload by filtering, categorizing, ranking, alerting, and monitoringdigital sources

• “Delivering the right content, to the right user, at the most adequate time, and in the most suitable way”

28/07/2005 27


B2B B2B -- B2B2CB2B2C(info producers, information brokers, info re-disributors, service&content providers,

vertical portals, infomediators, web clippers,…)

on PC, PDA, mobile devices(wap, gprs, umts,...), TV,

playstation,...

DigitalDigital EE--ContentContentSpaceSpace

infoinfoFACTORYFACTORYINTERNET

INTERNET /INTRANET

/INTRANET

(web sites,

portals, n

ews,

directories

, e-new

papers, e-

mail, web TV/rad

io,web-cams,

...)

DIGITAL TV

DIGITAL TV

(digitalTV broadcas

t, sate

llite

TV channels

, intera

ctive TV,

stream

ing service

s, video-

conferences

, ...)

text

text/data

, 2D/3D

/data, 2D/3D grap

hics

graphics,

, digitaldigital

audio/video,

audio/video, ……

(any format,

any standard

)REPOSITORIES

REPOSITORIES

(data banks, O

PAC, Digital

Libraries,

CD-Rom, ...)

Personalize

d

Personalize

d Information

Information Serv

ices

Service

s

EE--ContentContent ConsumerConsumerSpaceSpace

B2CB2C(individual end users, virtualcommunities, associations,

learneres, …)

-- sourcesource discoverydiscovery-- contentcontent identificationidentification-- contentcontent retrievalretrieval-- contentcontent filteringfiltering-- contentcontent classificationclassification-- contentcontent rankingranking-- content recontent re--mixingmixing-- c. recommendationc. recommendation-- content content reformulationreformulation-- contentcontent monitoringmonitoring-- contentcontent deliverydelivery-- content pushcontent push-- alertingalerting-- user profilinguser profiling-- conceptual conceptual analysisanalysis


28/07/2005 28


The infoinfoFACTORYFACTORY Technical Approach

• Cognitive (content-based) Filtering– Analysis of the conceptual content of textual documents– Semantic networks for describing user interests– Filtering obtained by matching documents against profiles

• Adaptive Personalization– Individual user profiles– Machine learning algorithms for automatically building

and updating user profiles on the basis of positive (negative) examples

28/07/2005 29


Intelligent Platforms

for achievingflexibility, re-usability, interoperability,

scalability, …

in providing personalized informationservices

28/07/2005 30


Cognitive filtering

Adaptive Personalization Classification Intelligent Crawling

Conceptual analysisof natural language

ifMONITOR ifEXPERT ifMAIL Others..

Contruction of personal Data Banks

Web Monitoring and Clippingservice

Finding Experts

Match mackingmodule betweenexpertcompetences and requests

E-mail filtering

Automaticcategorization

Antispam Filter

Multichannel Opp. Alerting

infoFACTORY technologies and basic components

28/07/2005 31


Cognitive filtering

Adaptive Personalization Classification Intelligent Crawling

Conceptual analysisof natural language

ifMONITOR ifEXPERT ifMAIL Others..

Contruction of personal Data Banks

Web Monitoring and Clippingservice

Finding Experts

Match mackingmodule betweenexpertcompetences and requests

E-mail filtering

Automaticcategorization

Antispam Filter

Multichannel Alerting

infoFACTORY technologies and basic components

Applicatio

n Level

System Level

PLATFORM Level

IntelligentIntelligent Platform1:Platform1:

ififMONITORMONITOR

forfor Web Web MonitoringMonitoring and and FilteringFiltering

28/07/2005 33


FilteringFiltering Web Web InformationInformation

• Using the IFT algorithm for filtering informationgathered from the Web by means of an opportunistic(automatic) human like navigation (crawling), relevantfor a given user profile

(Tasso and Armellini, 1999)

• Integration of various technologies and tools

28/07/2005 34


MonitoringMonitoring Web Web InformationInformation

• Periodically analysing the content of Web sites and portals for timely discovering new relevant information

• Search Engine monitoring

28/07/2005 35

infoinfoFACTORYFACTORYifMONITOR ifMONITOR ServicesServices

Personalized Filtering Services based on conceptual analisysof textual Web documents:

• Automatic construction of personalized THEMATIC DATA BANKS of Web documents

• WEB MONITORING, continuous periodic control of online sources– and multichannel ALERTING

28/07/2005 36


What are they useful for?What are they useful for?

• Technological monitoring• Monitoring scientific and technical journals• Reputation management• Press review• Competitive intelligence• Keeping knowledge worker always updated• Crisis management• Monitoring events• Monitoring financial funding open calls• Discovering research groups belonging to other

organizations in the same area of interests or research

28/07/2005 37


What advantages?What advantages?

• Reducing manual activities for searching information (higherproductivity – 24/24, 7/7 and 365/365)

• Finding information published in sites not (or rarely) indexed by standard search engines

• Timely discovery and multi-channel delivery of new relevantinformation

• Much more precise than other competitors (90% vs. 20-30% of standard search engines)

• Innovative information services (personalized services, alerting, mobile access)

UTENTE(Addetto Comunicazione, Ufficio Stampa, DG, …) WEB

Argomenti diinteresse,

esempi ififMonitorMonitorDocumentidell’utente

profili

Banca Dati TematicaPersonalizzata di doc. Web(che cresce incrementalmentenel tempo)

ififPortalPortal

Consultazionee ricerca

tramiteMotore

Inserimentodocumentidal proprioPC*

Costruzioneiniziale BD Monitoraggio

Periodico

nuove notizie

Inserimentomanuale dinotizie tramiteScanner oda Web*

PortaleIstituzionale

Pubblicazione Automatica*(previa validazione opt-in)

Sorgenti Online

Utility(ordinamenti,cartelle,pdf, mail,annotazione*,opt-in*)

ififClassifyClassify

Suddivisione insottocategorie

ed inoltro*

“Dicono di noi…”

Statistichesui media

* Servizi Opzionali

28/07/2005 39


ifPORTAL: a portal for accessingifMONITOR services

28/07/2005 49

infoinfoFACTORYFACTORYMobile access

28/07/2005 50


Some technical aspect

• Filtering Technology:– Very High precision by combining IFT content-

based filtering algorithm with key-word matchingand regular expressions

– High recall, with notification of possible faults– Innovation in the matching algorithm: conceptual

matching

28/07/2005 51


Conceptual Matching

Matching based on:• 1. Conceptual Coverage: how much the

most relevant conceptual content of the profile is covered in the evaluated doc.

• 2. Match: how much relevant concepts are present in the evaluated doc.

• Rank: a 3-star ranking obtained as a combination of the previous two measures

28/07/2005 52


Some technical aspect/2

• Construction of a personalized thematicData Bank:– Bulid the profile:

• Definition through: Key words, texts, URLs• Refinement through: relevance feedback

– Meta searching– Opportunistic navigation through hyperlinks

based on relevance of Web documents againstpersonal profile

28/07/2005 53


Opportunistic Crawling

28/07/2005 54


Notifying candidate faults

28/07/2005 55


IntelligentIntelligent PlatformPlatform 2:2:

ififEXPERTEXPERT

forfor ExpertExpert FindingFinding

28/07/2005 56


RepresentingRepresenting humanhuman expertise and expertise and searchingsearching forfor itit

• Using the same approach of the IFT algorithm for analysingtext and matching against profiles

• ‘Competence profiles’ are built (and later updated) byanalysing textual information describing personal expertise, projects, papers, and so on

• A ‘premium service’ of the MACSI-net Web site: match-making between requests for experts and an expert data base

• An ‘innovative service’ in the ‘Portale TIROCINI (Stage Portal)’, a portal for managing stages in industries foruniversity students: match-making between stage requestsand student competence & preferences.

28/07/2005 57


www.MACSINET.org:a WEB site exploiting the ifEXPERT

platform

28/07/2005 58


28/07/2005 59


Portale TIROCINI: a WEB portal exploiting the ifEXPERT

match-making capabilities

More than 36.000Students, with theirCompetences, credits,preferences

More than 1800Companies andPublic Agencies

Structuredinformation

Unstructuredinformation

ifEXPERTautomaticallyidentifiesthe bestmatchingprojects,taking intoaccountunstructuredinformation

28/07/2005 65



ififMAILMAIL

forfor Mail Mail FilteringFiltering and and AntispamAntispam

28/07/2005 66


CigniniCignini, M. , M. MizzaroMizzaro, S. and Tasso C. , S. and Tasso C. EE--mailmail CategorizationCategorization, , FilteringFiltering, and , and AlertingAlerting on Mobile on Mobile DevicesDevices: The : The ifMailifMail PrototypePrototype and and itsits ExperimentalExperimental

EvaluationEvaluation. In A. Cappelli and . In A. Cappelli and F.F. TuriniTurini ((EdsEds.) .) AI*IAAI*IA 2003: 2003: AdvancesAdvances in in ArtificialArtificial Intelligence Intelligence -- ProceedingProceeding of the 8th of the 8th CongressCongress of the of the ItalianItalian

AssociationAssociation forfor ArtificialArtificial IntelligenceIntelligence, Pisa, I, , Pisa, I, SeptemberSeptember 2003, 2003, SpringerSpringer VerlagVerlag, , BerlinBerlin, LNAI 2829, 2003, pp. 523, LNAI 2829, 2003, pp. 523--535.535.

CigniniCignini, M. , M. MizzaroMizzaro, S., Tasso C. and Virgili A. E, S., Tasso C. and Virgili A. E--mail on the Move: mail on the Move: Categorization, Filtering, and Alerting on Mobile Devices with tCategorization, Filtering, and Alerting on Mobile Devices with the ifMail he ifMail Prototype. In Prototype. In F.F. CrestaniCrestani, M. , M. DunlopDunlop and S. and S. MizzaroMizzaro ((EdsEds.) .) Mobile and Mobile and

UbiquitousUbiquitous InformationInformation Access Access -- Mobile HCI 2003 Mobile HCI 2003 InternationalInternational WorkshopWorkshop, , Udine, I, Udine, I, SeptemberSeptember 2003, 2003, SpringerSpringer VerlagVerlag, , BerlinBerlin, LNCS 2954, 2004, pp. 107, LNCS 2954, 2004, pp. 107--

123.123.

28/07/2005 67


FilteringFiltering and and ClassificationClassificationof eof e--mail mail MessagesMessages

• Filtering incoming e-mail messages, discarding the notinteresting ones (spam)

• Automatically categorizing messages into user-definedcategories of specific interest

• Category profiles are learned automatically, by observinguser’s behaviour

• Identifying the most relevant messages and alerting on Mobile devices

28/07/2005 69

infoinfoFACTORYFACTORYfor each message:

term extractioninternal representation

matching with profiles associated withcategories/folders:

categorizationfilteringalerting about most relevant

messages

profile construction, refinement, and evolution:

implicit relevance feedbackdacay function(rent)

stream of messages

User

Extraction of structured andunstructured data

internal representation

CategorizationModule

Categorized and FilteredMessages

C1 C2 CkC3

Dimensional reductionof unstructured terms

terms and weights

The user provides feedbackwhen he moves messagesamong folders

Filtering and Alerting Module

Mobile devices

Alerting

28/07/2005 70


Internal RepresentationContructor Module

CategorizationModule

InternalRepresentation

Database

Update state request

Monitoring activationrequest

Request or construction of aninternal representationInternal

representation

Feedback and Categorization request

New InternalRepresentation

Old InternalRepresentation

Answers

Categorization results

Feedback andCategorization request

Notification of intresting messages

Categorizedmessages

Mobiledevice

WebMail ModuleUser

E-mail Database

Commands

Results

Extract datarequest

Update datarequest

Mail Server(IMAP/SMTP)

Outpute-mails

Inpute-mails

Monitoring AgentModule

MultiChannelAlertingModule

Mail Filtering & Classification Engine

28/07/2005 71


0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1

15 45 75 105 135 165 195 225 255 285 315 345 375 405 435 465 495 525 555 585

Categorized e-mail

F1

F1 - 'Session' modality F1 - 'One-by-one' modality

moving average (Session) moving average (One-by-one)

Log. (moving average (One-by-one)) Log. (moving average (Session))

F1 in bothoperation modes for collection E

28/07/2005 72


0

10

20

30

40

50

60

0 100 200 300 400 500 600

Number of received messages

Cat

egor

ized

e-m

ails

System categorization User categorization

Log. (System categorization) Log. (User categorization)

Evaluation over timeof users’ vs. system actions(over groups of 60 messages):

i.e. how much effort is required from the user

28/07/2005 73



mymyTWMTWM

EE--LearningLearning PlatformPlatform forfor delivering delivering valuevalue--added services to University added services to University

studentsstudents

28/07/2005 74


PersonalizationPersonalization in the TWM in the TWM PortalPortal forfor valuevalue--addedadded servicesservices toto University University studentsstudents

• The TWM portal provides several services for University students

• Several services are supporting learning activities

• The student can personalize the portal and access onlyto the preferred services, and associate alerting

• Fully XML-SCORM compliant

28/07/2005 75


28/07/2005 76


28/07/2005 77


ifFORUM: an advanced forum,

with personalization andtools for anaysing social interaction

in e-learningand

knowledge management

Standard Forum Standard Forum CapabilitiesCapabilities

MessageMessage exchangeexchange networknetwork

CentralityCentrality measuremeasure: : discoveringdiscovering ‘‘leadersleaders’’

PCA and PCA and cliqueclique analysisanalysis: : discoveringdiscovering groupsgroups

DiscoveringDiscovering usersusers thatthat are are ‘‘informationinformation bridgesbridges’’

ExploitingExploitingpersonalizationpersonalizationandandfilteringfilteringforfor proactiveproactiveRecommendationRecommendation

in in ee--learninglearningandandknowledgeknowledgemanagementmanagement

Tasso, C., Rossi, P.G., Virgili, C., and Morandini, A., Content-based personalization for asynchronous communication tools: the ifFORUMsystem.

In L. Aroyo and C. Tasso (Eds.) AH 2004: 3rd International Conference on Adaptive Hypermedia and Adaptive Web-Based Systems – WorkshopProceedings, Technische Universiteit Eindhoven, Eindhoven, August 23rd, 2004, ISSN 0926-4515, pp. 324-330.

28/07/2005 84


CurrentCurrent ResearchResearch

• Personalization in Knowledge Management and e-Learning systems

• Social search, collaborative filtering (Districts of SMEs)

• ‘Personalizing’ the Medicine Library (BIBLIOMED Project)

28/07/2005 85


tÇ |Ç|à|tà|äx Éy à{x TÜà|y|v|tÄ \ÇàxÄÄ|zxÇvx _tuÉÜtàÉÜç?hÇ|äxÜá|àç Éy hw|Çx? ÑÜÉyA VtÜÄÉ gtááÉ

ãããA|ÇyÉytvàÉÜçA|à